V6 comparison to V5.x [solved]
Hi,
first - thank you very much for your work.
Hope it is okay, to start here a disussion / feedback.
My daily runner is the V55 since his release and i tested the V6 now many hours and
couldn get this realism as there was since your famous Euclid V5.0.
The skin texture and skin color on the V6 is (on my setup) very "render-ish". It is a little
better with your recommended Heun Sampler, but Heun is also a very "expensive" Sampler,
as it needs 2-3 the sample / diffusion time, also with lowered Steps.
V5.5 - out of the box - no VAE added - with many different samplers has a very realistic skin color
(red-elements on faces or knees etc.) and the wrinkles, skin texture can easy prompt in ..
My main issues are often the teeths, and on 5.1 there were some nose-ring issues. These are
much better on V5.5
May you have a final idea how to get (easy - also with Karras samplers) the nice skin colors / texture
back.
Again, thank you for your work
Hi Jott,
Thanks for taking the time to start this discussion, I'm not sure why you are getting render-ish skin on V6. I'll certainly look into this to see if it can be fixed. In the meantime would you be able to post an example prompt you use so I can test please.
Hi ConsistentFactor,
thank you for your answer and the option to send you a Prompt-Set. As those prompts
are actually "not for public" (... 4+ months of development and optimization - the
goal is to work with lower than 75 pos / 75 neg prompts and no Extra-VAE, Lora, TI, etc. with
absolut reality (sfw & nsfw) and full control of 12+ content-rules , in a runtime under
1 sec. / img on 4090s), i tried to find any possibility to send those per DM. But i couldnt
find here or in civtai any DM options. May you have a Discord or a other DM option?
After some re-work and much more testing on the V6 my results are nearly perfect.
The mayor culprit was my longtime used "semirealistic" as the first prompt inside my
negative chain. It really needs that "unrealistic" (as in your examples written) prompt (best on first place)
to diffuse the right direction.Also my "background and lighting" prompts should be on the (nearly)
last position in the positive prompts, to keep a consistent saturation in batch-runs. ..
I also get more friendship with the Heun Sampler. Lowering to 18-20 and little uplift of the cfg around 9 - 9.5
is enough and okay from the speed.
A nice Sampler, on lowerg cfg (7) and lower Sample Steps (20) is the DPM++ 2M SDE (not Karras)
It brings on those low - and very fast runs - some nice structure and grain to my results.
DDIM brings on Stepsizes around 10 great results. ( all tests are done without attention optimzations / xformers).
My Sampler summary:
(F -> Fast , S -> Slow, * = best results (subjectiv) )
F Euler a
F Euler
F LMS
S Heun *
S DPM2
S DPM2 a
S DPM++ 2S a
F DPM++ 2m *
S DPM++ SDE
F DPM++ 2M SDE *
F DPM fast
S DPM adaptive
F LMS Karras *
S DPM2 Karras *
S DPM2 a Karras *
S DPM++ 2S a Karras *
F DPM++ 2M Karras *
S DPM++ SDE Karras
F DPM++ 2M SDE Karras *
F DDIM *
F PLMS
F UNIPC
Thanks.
I'm glad you managed to solve it. If you have any questions feel free to contact me on discord, I'm not overly active but do check daily. My discord handle is ConsistentFactor444#2108
With regards to sampler testing I will do some more in depth testing. Will definitely try "Lowering to 18-20 and little uplift of the cfg around 9 - 9.5
is enough and okay from the speed."
Would help me as Heun only gets around 2it/s and then 2s/it on hi-res fix, at least on my hardware.
May you can also check with the .. DDIM and low Steps .. (tested on your uploaded V6 cute austrian .. png)
(4090 no attention opt. ):
DDIM - 12 steps , cfg 7 : 4.29 it/s -- vs -- Heun -35 steps , cfg 7 : 1.89it/s
(4090 WITH attention opt.):
DDIM - 12 steps , cfg 7 : 16.29 it/s -- vs -- Heun -35 steps , cfg 7 : 6.05it/s